Claude Architect · 课时

子代理不会继承历史记录

在每个子代理的提示词中显式传递所有必要的上下文。

子代理不会继承历史记录 是 CoddyKit 上的免费 Claude Architect 课时。这是第 3 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 Claude Architect 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Claude Architect 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

The Memory Trap

You build a hub-and-spoke multi-agent system. The coordinator has had a long conversation with the user: requirements, constraints, prior decisions. Then it delegates a task to a subagent and assumes the subagent already "knows" all of that.

It does not. This is the single most common multi-agent bug, and it appears directly on the Claude Certified Architect exam.

The rule: subagents do NOT inherit the coordinator's conversation history. Every piece of context a subagent needs must be passed explicitly in its prompt.

Why History Doesn't Transfer

The Claude API is stateless. The model keeps NO server-side memory between requests. On every turn you resend the FULL messages history yourself.

A subagent runs as its own independent loop, with its own messages array. The coordinator's history lives in the coordinator's request, not in some shared global memory. Nothing copies it across.

So when a subagent starts, its context is exactly what you put in its system prompt and first messages entry — and nothing more.

# Each agent owns its own messages array.
# Nothing is shared automatically between them.
resp = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=2048,
    system=subagent_system_prompt,   # subagent's OWN instructions
    messages=subagent_messages,      # subagent's OWN history (starts empty)
)

Hub-and-Spoke, Restated

In the hub-and-spoke pattern the coordinator decomposes the work, delegates to subagents, then aggregates and routes the results. It owns orchestration and error handling.

But delegation is a one-way handoff of explicit instructions, not a shared brain. Think of each subagent as a brand-new contractor who has never seen your earlier emails. You must brief them fully in the work order itself.

Coordinator: decompose, delegate, aggregate, route, handle errors.
Subagent: receives a self-contained brief, does one focused job, returns a result.

The Failure Mode

Here is what the bug looks like in practice. The user told the coordinator the target is the checkout service and the deadline is strict. The coordinator then spawns a reviewer subagent with a vague brief.

The subagent has no idea which service, which constraints, or what "the file" refers to. It hallucinates a target, reviews the wrong thing, or asks a question the coordinator can't relay back cleanly.

# ANTI-PATTERN: assumes the subagent 'remembers' the chat
Task(
    description="Review the file",
    prompt="Review the file we discussed and flag any bugs.",
    subagent_type="code-reviewer",
)
# 'the file we discussed' means NOTHING to a fresh subagent.

The Fix: Self-Contained Briefs

Rewrite the brief so it stands entirely on its own. Pass the target, the constraints, the prior decisions, and the exact output you expect.

A good subagent prompt answers: What is the task? On what exact input? Under what constraints? In what output format? No reference to "earlier" or "as we said" survives the handoff.

Task(
    description="Review checkout/payment.py",
    prompt=(
        "Review the file checkout/payment.py for correctness bugs.\n"
        "Context: this handles card charges; idempotency is required.\n"
        "Prior decision: refunds over $500 must route to a human.\n"
        "Flag a comment ONLY when it contradicts the code.\n"
        "Return findings as a JSON list of {line, severity, issue}."
    ),
    subagent_type="code-reviewer",
)

Pass Facts Verbatim, Not Vibes

Don't summarize transactional facts into mush before handing them off. Progressive summarization makes numbers, percentages, and dates vague — and a subagent acting on "roughly last quarter" instead of "2026-Q1" will be wrong.

Keep the hard facts a subagent needs in a verbatim "context" block: IDs, thresholds, file paths, exact dates. Summarize prose for flavor; never summarize the load-bearing details.

context_block = (
    "CASE FACTS (verbatim):\n"
    "- customer_id: CUS-88231 (identity verified)\n"
    "- order_id: ORD-55012\n"
    "- refund_amount: $512.00  (exceeds $500 policy threshold)\n"
    "- requested_date: 2026-06-10\n"
)
subagent_prompt = context_block + "\nTask: draft the refund-approval request."

Least Privilege Travels With the Brief

An AgentDefinition carries: name, description, system_prompt, and allowed_tools. Because the subagent is isolated, its tools and its system prompt ARE its whole world — scope them to the role.

Give a subagent the 4-5 tools it actually needs (4-5 per agent is optimal; 18+ degrades tool selection). And remember: for the coordinator to spawn subagents at all, the coordinator's allowedTools must include "Task".

reviewer = AgentDefinition(
    name="code-reviewer",
    description="Reviews one file for correctness bugs; returns JSON findings.",
    system_prompt=(
        "You review a single file. All needed context is in the user "
        "message. Never assume prior conversation exists."
    ),
    allowed_tools=["Read", "Grep"],   # least privilege
)

Parallel Subagents Are Fully Independent

Issuing multiple Task calls in a single response runs them in parallel. That is powerful — but it doubles down on the isolation rule.

Parallel subagents cannot see each other's history OR the coordinator's. Each must be briefed independently and completely. There is no implicit ordering and no shared scratchpad between them; if subagent B needs subagent A's output, the coordinator must collect A's result and feed it into B's prompt explicitly.

# Two Task calls in ONE response -> run in parallel, fully isolated.
# Each gets its OWN complete brief; neither sees the other's.
Task(prompt=brief_for_auth_module, subagent_type="code-reviewer")
Task(prompt=brief_for_billing_module, subagent_type="code-reviewer")

Returning Results: Structured, Not Chatty

Isolation also shapes the return trip. The coordinator only gets back what the subagent emits — so make that emission machine-usable.

For aggregation, force structured output: a subagent with tool_choice="any" MUST call a tool, which guarantees the coordinator receives parseable JSON instead of free prose it has to scrape. And when a subagent fails, it should return structured context (failure type, attempted query, partial results) — not a generic "operation failed" that blocks recovery.

resp = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    system=subagent_system_prompt,
    messages=subagent_messages,
    tools=[report_findings_tool],
    tool_choice={"type": "any"},   # guarantees a structured result back to the hub
)

Sessions Are Not a Loophole

You might hope a resumed session smuggles history into a subagent. Be careful. --resume <name> continues a named session and fork_session branches from a shared point — but these resume a session's own state, they do not retroactively inject the coordinator's chat into a fresh subagent.

And resumed tool results can be stale if the codebase changed underneath them. Sometimes a fresh session seeded with a structured summary beats resuming — which is exactly the explicit-context discipline again.

A Practical Briefing Checklist

Before you spawn any subagent, confirm its prompt is self-contained. Walk this checklist:

Target: the exact file / record / id it operates on.
Constraints: policies, thresholds, prior decisions — verbatim.
Task: one focused job, stated with explicit criteria.
Output: the exact shape to return (JSON schema / fields).
Tools: only the 4-5 it needs, least privilege.

If you removed the coordinator entirely and handed this prompt to a stranger, could they do the job? If yes, you've briefed it correctly.

Quick Check: The Forgetful Subagent

A coordinator has spent 20 turns clarifying that the user wants a security review of auth/session.py, with the rule "only flag findings that are exploitable in production." It now delegates to a reviewer subagent. What is the correct way to delegate?

Recap: Brief Every Subagent Fully

Key takeaways:

The API is stateless and subagents are isolated — they inherit none of the coordinator's history.
Every subagent prompt must be self-contained: target, constraints, task, output format, tools.
Pass transactional facts (ids, thresholds, dates, paths) verbatim; don't let summarization blur them.
Parallel Task calls are independent — brief each one separately; the coordinator feeds one subagent's output into another explicitly.
Scope allowed_tools to the role (4-5 optimal); the coordinator needs "Task" to delegate.
Use tool_choice="any" for structured returns and structured errors for recoverable aggregation.

Brief the stranger, not the friend. That mindset passes both the exam and production.

常见问题解答

「子代理不会继承历史记录」课时是免费的吗？

是的 — 「子代理不会继承历史记录」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Claude Architect 课程的其余内容，请升级到 CoddyKit PRO。 Claude Architect 课程共包含 4 节课。

「子代理不会继承历史记录」这节课中我会学到什么？

在每个子代理的提示词中显式传递所有必要的上下文。你通过在浏览器中直接运行的动手代码来练习 Claude Architect，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 Claude Architect 需要有经验吗？

无需任何先前经验。CoddyKit 上的 Claude Architect 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 3 节课，共 4 节。

「子代理不会继承历史记录」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 Claude Architect 课中编写并运行代码吗？

能。每节 Claude Architect 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 Claude Architect